OpenAI has introduced a new sandbox environment for its Codex coding agent on Windows, designed to improve both safety and usability. The company previously lacked a sandbox implementation for Windows, forcing users to choose between excessive command approvals or unrestricted access. The new sandbox enforces strict boundaries for file writes and network access, ensuring that Codex operates safely within user workspaces while maintaining productivity.

The development followed an evaluation of existing Windows isolation tools such as AppContainer, Windows Sandbox, and Mandatory Integrity Control labeling, all of which proved unsuitable for Codex’s open-ended workflows. OpenAI first built an unelevated prototype using Windows security identifiers (SIDs) and write-restricted tokens to control file access, but this version offered only weak network protection. To address that limitation, the team created an elevated sandbox that uses dedicated Windows user accounts and firewall rules to enforce stronger network restrictions.

The elevated sandbox requires administrative setup but provides more robust isolation and control. This implementation aims to make Codex as secure and efficient on Windows as it is on macOS and Linux, offering developers a safer environment for AI-assisted coding.